X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_io.c;h=85e8310e536048e0682f317ada0faa32f4ae2d48;hp=e0b07485ec15add72f05a0400d8429ba3e78b3ef;hb=0754bc8f2623bea184111af216f7567608db35b6;hpb=6fc066434a3b8d4cfee0e038eb5da1cf4c8b5c74 diff --git a/lustre/osd-zfs/osd_io.c b/lustre/osd-zfs/osd_io.c index e0b0748..85e8310 100644 --- a/lustre/osd-zfs/osd_io.c +++ b/lustre/osd-zfs/osd_io.c @@ -26,10 +26,8 @@ /* * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. - */ -/* - * Copyright (c) 2012, 2013, Intel Corporation. - * Use is subject to license terms. + * + * Copyright (c) 2012, 2015, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -51,7 +49,7 @@ #include #include #include -#include /* LLOG_CHUNK_SIZE definition */ +#include /* LLOG_MIN_CHUNK_SIZE definition */ #include "osd_internal.h" @@ -70,19 +68,64 @@ static char *osd_zerocopy_tag = "zerocopy"; + +static void record_start_io(struct osd_device *osd, int rw, int discont_pages) +{ + struct obd_histogram *h = osd->od_brw_stats.hist; + + if (rw == READ) { + atomic_inc(&osd->od_r_in_flight); + lprocfs_oh_tally(&h[BRW_R_RPC_HIST], + atomic_read(&osd->od_r_in_flight)); + lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages); + + } else { + atomic_inc(&osd->od_w_in_flight); + lprocfs_oh_tally(&h[BRW_W_RPC_HIST], + atomic_read(&osd->od_w_in_flight)); + lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages); + + } +} + +static void record_end_io(struct osd_device *osd, int rw, + unsigned long elapsed, int disksize, int npages) +{ + struct obd_histogram *h = osd->od_brw_stats.hist; + + if (rw == READ) { + atomic_dec(&osd->od_r_in_flight); + lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages); + if (disksize > 0) + lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize); + if (elapsed) + lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed); + + } else { + atomic_dec(&osd->od_w_in_flight); + lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages); + if (disksize > 0) + lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize); + if (elapsed) + lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed); + } +} + static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt, - struct lu_buf *buf, loff_t *pos, - struct lustre_capa *capa) + struct lu_buf *buf, loff_t *pos) { struct osd_object *obj = osd_dt_obj(dt); struct osd_device *osd = osd_obj2dev(obj); uint64_t old_size; int size = buf->lb_len; int rc; + unsigned long start; LASSERT(dt_object_exists(dt)); LASSERT(obj->oo_db); + start = cfs_time_current(); + read_lock(&obj->oo_attr_lock); old_size = obj->oo_attr.la_size; read_unlock(&obj->oo_attr_lock); @@ -94,16 +137,16 @@ static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt, size = old_size - *pos; } - rc = -dmu_read(osd->od_objset.os, obj->oo_db->db_object, *pos, size, + record_start_io(osd, READ, 0); + + rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size, buf->lb_buf, DMU_READ_PREFETCH); + + record_end_io(osd, READ, cfs_time_current() - start, size, + size >> PAGE_CACHE_SHIFT); if (rc == 0) { rc = size; *pos += size; - - /* XXX: workaround for bug in HEAD: fsfilt_ldiskfs_read() returns - * requested number of bytes, not actually read ones */ - if (S_ISLNK(obj->oo_dt.do_lu.lo_header->loh_attr)) - rc = buf->lb_len; } return rc; } @@ -140,9 +183,9 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, /* XXX: we still miss for append declaration support in ZFS * -1 means append which is used by llog mostly, llog - * can grow upto LLOG_CHUNK_SIZE*8 records */ + * can grow upto LLOG_MIN_CHUNK_SIZE*8 records */ if (pos == -1) - pos = max_t(loff_t, 256 * 8 * LLOG_CHUNK_SIZE, + pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE, obj->oo_attr.la_size + (2 << 20)); dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len); @@ -156,15 +199,14 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, const struct lu_buf *buf, loff_t *pos, - struct thandle *th, struct lustre_capa *capa, - int ignore_quota) + struct thandle *th, int ignore_quota) { struct osd_object *obj = osd_dt_obj(dt); struct osd_device *osd = osd_obj2dev(obj); - udmu_objset_t *uos = &osd->od_objset; struct osd_thandle *oh; uint64_t offset = *pos; int rc; + ENTRY; LASSERT(dt_object_exists(dt)); @@ -173,7 +215,9 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, LASSERT(th != NULL); oh = container_of0(th, struct osd_thandle, ot_super); - dmu_write(osd->od_objset.os, obj->oo_db->db_object, offset, + record_start_io(osd, WRITE, 0); + + dmu_write(osd->od_os, obj->oo_db->db_object, offset, (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx); write_lock(&obj->oo_attr_lock); if (obj->oo_attr.la_size < offset + buf->lb_len) { @@ -182,7 +226,7 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, /* osd_object_sa_update() will be copying directly from oo_attr * into dbuf. any update within a single txg will copy the * most actual */ - rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos), + rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd), &obj->oo_attr.la_size, 8, oh); if (unlikely(rc)) GOTO(out, rc); @@ -194,6 +238,9 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, rc = buf->lb_len; out: + record_end_io(osd, WRITE, 0, buf->lb_len, + buf->lb_len >> PAGE_CACHE_SHIFT); + RETURN(rc); } @@ -201,8 +248,8 @@ out: * XXX: for the moment I don't want to use lnb_flags for osd-internal * purposes as it's not very well defined ... * instead I use the lowest bit of the address so that: - * arc buffer: .lnb_obj = abuf (arc we loan for write) - * dbuf buffer: .lnb_obj = dbuf | 1 (dbuf we get for read) + * arc buffer: .lnb_data = abuf (arc we loan for write) + * dbuf buffer: .lnb_data = dbuf | 1 (dbuf we get for read) * copy buffer: .lnb_page->mapping = obj (page we allocate for write) * * bzzz, to blame @@ -253,14 +300,37 @@ static inline struct page *kmem_to_page(void *addr) return virt_to_page(addr); } +/** + * Prepare buffers for read. + * + * The function maps the range described by \a off and \a len to \a lnb array. + * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then + * we fill \a lnb array with the pages storing ARC buffers. Notice the current + * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC + * buffers with actual data, I/O is done in the conext of osd_bufs_get_read(). + * A better implementation would just return the buffers (potentially unfilled) + * and subsequent osd_read_prep() would do I/O for many ranges concurrently. + * + * \param[in] env environment + * \param[in] obj object + * \param[in] off offset in bytes + * \param[in] len the number of bytes to access + * \param[out] lnb array of local niobufs pointing to the buffers with data + * + * \retval 0 for success + * \retval negative error number of failure + */ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj, loff_t off, ssize_t len, struct niobuf_local *lnb) { struct osd_device *osd = osd_obj2dev(obj); - dmu_buf_t **dbp; + unsigned long start = cfs_time_current(); int rc, i, numbufs, npages = 0; + dmu_buf_t **dbp; ENTRY; + record_start_io(osd, READ, 0); + /* grab buffers for read: * OSD API let us to grab buffers first, then initiate IO(s) * so that all required IOs will be done in parallel, but at the @@ -297,12 +367,12 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj, lnb->lnb_rc = 0; lnb->lnb_file_offset = off; - lnb->lnb_page_offset = bufoff & ~CFS_PAGE_MASK; + lnb->lnb_page_offset = bufoff & ~PAGE_MASK; lnb->lnb_len = thispage; lnb->lnb_page = kmem_to_page(dbp[i]->db_data + bufoff); /* mark just a single slot: we need this - * reference to dbuf to be release once */ + * reference to dbuf to be released once */ lnb->lnb_data = dbf; dbf = NULL; @@ -315,13 +385,17 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj, lnb++; } - /* steal dbuf so dmu_buf_rele_array() cant release it */ + /* steal dbuf so dmu_buf_rele_array() can't release + * it */ dbp[i] = NULL; } dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag); } + record_end_io(osd, READ, cfs_time_current() - start, + npages * PAGE_SIZE, npages); + RETURN(npages); err: @@ -436,7 +510,7 @@ out_err: static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt, loff_t offset, ssize_t len, struct niobuf_local *lnb, - int rw, struct lustre_capa *capa) + int rw) { struct osd_object *obj = osd_dt_obj(dt); int rc; @@ -463,86 +537,49 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, return 0; } -/* Return number of blocks that aren't mapped in the [start, start + size] - * region */ -static int osd_count_not_mapped(struct osd_object *obj, uint64_t start, - uint32_t size) +static inline uint32_t osd_get_blocksz(struct osd_object *obj) { - dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)obj->oo_db; - dmu_buf_impl_t *db; - dnode_t *dn; - uint32_t blkshift; - uint64_t end, blkid; - int rc; - ENTRY; - - DB_DNODE_ENTER(dbi); - dn = DB_DNODE(dbi); - - if (dn->dn_maxblkid == 0) { - if (start + size <= dn->dn_datablksz) - GOTO(out, size = 0); - if (start < dn->dn_datablksz) - start = dn->dn_datablksz; - /* assume largest block size */ - blkshift = SPA_MAXBLOCKSHIFT; - } else { - /* blocksize can't change */ - blkshift = dn->dn_datablkshift; - } + uint32_t blksz; + u_longlong_t unused; - /* compute address of last block */ - end = (start + size - 1) >> blkshift; - /* align start on block boundaries */ - start >>= blkshift; + LASSERT(obj->oo_db); - /* size is null, can't be mapped */ - if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0) - GOTO(out, size = (end - start + 1) << blkshift); + dmu_object_size_from_db(obj->oo_db, &blksz, &unused); + return blksz; +} - /* beyond EOF, can't be mapped */ - if (start > dn->dn_maxblkid) - GOTO(out, size = (end - start + 1) << blkshift); +static inline uint64_t osd_roundup2blocksz(uint64_t size, + uint64_t offset, + uint32_t blksz) +{ + LASSERT(blksz > 0); - size = 0; - for (blkid = start; blkid <= end; blkid++) { - if (blkid == dn->dn_maxblkid) - /* this one is mapped for sure */ - continue; - if (blkid > dn->dn_maxblkid) { - size += (end - blkid + 1) << blkshift; - GOTO(out, size); - } + size += offset % blksz; - rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db); - if (rc) { - /* for ENOENT (block not mapped) and any other errors, - * assume the block isn't mapped */ - size += 1 << blkshift; - continue; - } - dbuf_rele(db, FTAG); - } + if (likely(IS_PO2(blksz))) + return PO2_ROUNDUP_TYPED(size, blksz, uint64_t); - GOTO(out, size); -out: - DB_DNODE_EXIT(dbi); - return size; + size += blksz - 1; + do_div(size, blksz); + return size * blksz; } static int osd_declare_write_commit(const struct lu_env *env, - struct dt_object *dt, - struct niobuf_local *lnb, int npages, - struct thandle *th) + struct dt_object *dt, + struct niobuf_local *lnb, int npages, + struct thandle *th) { struct osd_object *obj = osd_dt_obj(dt); struct osd_device *osd = osd_obj2dev(obj); struct osd_thandle *oh; uint64_t offset = 0; uint32_t size = 0; + uint32_t blksz = osd_get_blocksz(obj); int i, rc, flags = 0; bool ignore_quota = false, synced = false; long long space = 0; + struct page *last_page = NULL; + unsigned long discont_pages = 0; ENTRY; LASSERT(dt_object_exists(dt)); @@ -554,6 +591,9 @@ static int osd_declare_write_commit(const struct lu_env *env, oh = container_of0(th, struct osd_thandle, ot_super); for (i = 0; i < npages; i++) { + if (last_page && lnb[i].lnb_page->index != (last_page->index + 1)) + ++discont_pages; + last_page = lnb[i].lnb_page; if (lnb[i].lnb_rc) /* ENOSPC, network RPC error, etc. * We don't want to book space for pages which will be @@ -587,11 +627,13 @@ static int osd_declare_write_commit(const struct lu_env *env, dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset, size); - /* estimating space that will be consumed by a write is rather + /* Estimating space to be consumed by a write is rather * complicated with ZFS. As a consequence, we don't account for - * indirect blocks and quota overrun will be adjusted once the - * operation is committed, if required. */ - space += osd_count_not_mapped(obj, offset, size); + * indirect blocks and just use as a rough estimate the worse + * case where the old space is being held by a snapshot. Quota + * overrun will be adjusted once the operation is committed, if + * required. */ + space += osd_roundup2blocksz(size, offset, blksz); offset = lnb[i].lnb_file_offset; size = lnb[i].lnb_len; @@ -600,7 +642,7 @@ static int osd_declare_write_commit(const struct lu_env *env, if (size) { dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset, size); - space += osd_count_not_mapped(obj, offset, size); + space += osd_roundup2blocksz(size, offset, blksz); } dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0); @@ -609,11 +651,12 @@ static int osd_declare_write_commit(const struct lu_env *env, /* backend zfs filesystem might be configured to store multiple data * copies */ - space *= osd->od_objset.os->os_copies; + space *= osd->od_os->os_copies; space = toqb(space); - CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota " - "space\n", npages, space); + CDEBUG(D_QUOTA, "writing %d pages, reserving "LPD64"K of quota space\n", + npages, space); + record_start_io(osd, WRITE, discont_pages); retry: /* acquire quota space if needed */ rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid, @@ -639,16 +682,61 @@ retry: RETURN(rc); } +/** + * Policy to grow ZFS block size by write pattern. + * For sequential write, it grows block size gradually until it reaches the + * maximum blocksize the dataset can support. Otherwise, it will just use + * the maximum block size. + */ +static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh, + uint64_t start, uint64_t end) +{ + struct osd_device *osd = osd_obj2dev(obj); + dmu_buf_impl_t *db = (dmu_buf_impl_t *)obj->oo_db; + dnode_t *dn; + uint32_t blksz; + int rc = 0; + ENTRY; + + DB_DNODE_ENTER(db); + dn = DB_DNODE(db); + + if (dn->dn_maxblkid > 0) /* can't change block size */ + GOTO(out, rc); + + blksz = dn->dn_datablksz; + if (blksz >= osd->od_max_blksz) + GOTO(out, rc); + + /* now ZFS can support up to 16MB block size, and if the write + * is sequential, it just increases the block size gradually */ + if (start <= blksz) { /* sequential */ + blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end); + if (!is_power_of_2(blksz)) + blksz = size_roundup_power2(blksz); + } else { /* otherwise, use maximum block size */ + blksz = osd->od_max_blksz; + } + + if (blksz > dn->dn_datablksz) + rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object, + blksz, 0, oh->ot_tx); + EXIT; +out: + DB_DNODE_EXIT(db); + return rc; +} + static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, struct niobuf_local *lnb, int npages, struct thandle *th) { struct osd_object *obj = osd_dt_obj(dt); struct osd_device *osd = osd_obj2dev(obj); - udmu_objset_t *uos = &osd->od_objset; struct osd_thandle *oh; uint64_t new_size = 0; int i, rc = 0; + unsigned long iosize = 0; ENTRY; LASSERT(dt_object_exists(dt)); @@ -657,6 +745,14 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, LASSERT(th != NULL); oh = container_of0(th, struct osd_thandle, ot_super); + /* adjust block size. Assume the buffers are sorted. */ + rc = osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset, + lnb[npages - 1].lnb_file_offset + + lnb[npages - 1].lnb_len); + if (rc < 0) /* ignore the error */ + CDEBUG(D_INODE, "obj "DFID": change block size error rc=%d\n", + PFID(lu_object_fid(&dt->do_lu)), rc); + for (i = 0; i < npages; i++) { CDEBUG(D_INODE, "write %u bytes at %u\n", (unsigned) lnb[i].lnb_len, @@ -673,7 +769,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, } if (lnb[i].lnb_page->mapping == (void *)obj) { - dmu_write(osd->od_objset.os, obj->oo_db->db_object, + dmu_write(osd->od_os, obj->oo_db->db_object, lnb[i].lnb_file_offset, lnb[i].lnb_len, kmap(lnb[i].lnb_page), oh->ot_tx); kunmap(lnb[i].lnb_page); @@ -693,6 +789,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len) new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len; + iosize += lnb[i].lnb_len; } if (unlikely(new_size == 0)) { @@ -700,6 +797,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, th->th_local = 1; /* it is important to return 0 even when all lnb_rc == -ENOSPC * since ofd_commitrw_write() retries several times on ENOSPC */ + record_end_io(osd, WRITE, 0, 0, 0); RETURN(0); } @@ -710,12 +808,14 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, /* osd_object_sa_update() will be copying directly from * oo_attr into dbuf. any update within a single txg will copy * the most actual */ - rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos), - &obj->oo_attr.la_size, 8, oh); + rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd), + &obj->oo_attr.la_size, 8, oh); } else { write_unlock(&obj->oo_attr_lock); } + record_end_io(osd, WRITE, 0, iosize, npages); + RETURN(rc); } @@ -723,25 +823,28 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, struct niobuf_local *lnb, int npages) { struct osd_object *obj = osd_dt_obj(dt); - struct lu_buf buf; - loff_t offset; int i; + loff_t eof; LASSERT(dt_object_exists(dt)); LASSERT(obj->oo_db); + read_lock(&obj->oo_attr_lock); + eof = obj->oo_attr.la_size; + read_unlock(&obj->oo_attr_lock); + for (i = 0; i < npages; i++) { - buf.lb_buf = kmap(lnb[i].lnb_page); - buf.lb_len = lnb[i].lnb_len; - offset = lnb[i].lnb_file_offset; + if (unlikely(lnb[i].lnb_rc < 0)) + continue; - CDEBUG(D_OTHER, "read %u bytes at %u\n", - (unsigned) lnb[i].lnb_len, - (unsigned) lnb[i].lnb_file_offset); - lnb[i].lnb_rc = osd_read(env, dt, &buf, &offset, NULL); - kunmap(lnb[i].lnb_page); + lnb[i].lnb_rc = lnb[i].lnb_len; + + if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) { + if (eof <= lnb[i].lnb_file_offset) + lnb[i].lnb_rc = 0; + else + lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset; - if (lnb[i].lnb_rc < buf.lb_len) { /* all subsequent rc should be 0 */ while (++i < npages) lnb[i].lnb_rc = 0; @@ -787,12 +890,10 @@ static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx, } static int osd_punch(const struct lu_env *env, struct dt_object *dt, - __u64 start, __u64 end, struct thandle *th, - struct lustre_capa *capa) + __u64 start, __u64 end, struct thandle *th) { struct osd_object *obj = osd_dt_obj(dt); struct osd_device *osd = osd_obj2dev(obj); - udmu_objset_t *uos = &osd->od_objset; struct osd_thandle *oh; __u64 len; int rc = 0; @@ -812,15 +913,15 @@ static int osd_punch(const struct lu_env *env, struct dt_object *dt, len = end - start; write_unlock(&obj->oo_attr_lock); - rc = __osd_object_punch(osd->od_objset.os, obj->oo_db, oh->ot_tx, + rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx, obj->oo_attr.la_size, start, len); /* set new size */ if (len == DMU_OBJECT_END) { write_lock(&obj->oo_attr_lock); obj->oo_attr.la_size = start; write_unlock(&obj->oo_attr_lock); - rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos), - &obj->oo_attr.la_size, 8, oh); + rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd), + &obj->oo_attr.la_size, 8, oh); } RETURN(rc); } @@ -872,4 +973,3 @@ struct dt_body_operations osd_body_ops = { .dbo_declare_punch = osd_declare_punch, .dbo_punch = osd_punch, }; -